Google Releases Sashiko: AI Code Review Agent for Linux Kernel Patches
Google engineers have released Sashiko, an open-source agentic AI code review system specifically designed for the Linux kernel. Roman Gushchin of Google's Linux kernel team announced the project, which has been used internally at Google to uncover issues and is now publicly available.
Key Details
Sashiko covers all submissions to the Linux kernel mailing list and is funded by Google, which also covers the token budget and infrastructure. The project hosting is being moved to the Linux Foundation.
According to Roman's announcement, Sashiko was able to find around 53% of bugs based on an unfiltered set of 1,000 recent upstream Linux kernel issues with "Fixes:" tags using Gemini 3.1 Pro. He noted that "100% of these issues were missed by human reviewers."
The system was designed for Google Gemini Pro 3.1 but should work with Claude and likely other LLMs too.
Access and Resources
- The code is open-source on GitHub
- A web interface for upstream agentic code review of Linux kernel patches is available at Sashiko.dev
- Roman's full announcement can be read on LinkedIn
This tool is for developers working on or reviewing Linux kernel code, particularly those submitting patches to the upstream mailing list.
📖 Read the full source: HN AI Agents
👀 See Also
Cloudflare Dynamic Worker Loader: Sandboxing AI Agents with Isolates
Cloudflare's Dynamic Worker Loader API, now in open beta, allows Workers to instantiate new Workers with runtime-specified code in isolated sandboxes using V8 isolates, offering 100x faster startup than containers and no global concurrency limits.
Skales Desktop AI Agent Built with Claude, Features Clippy-Style Mascot
Skales is a desktop AI agent that runs locally on Windows and macOS, using Claude via OpenRouter/Anthropic API for reasoning and tool execution. It includes a floating Desktop Buddy mascot with a paperclip skin reference and can execute commands like sending emails, managing files, browsing the web, and managing calendars.
OpenClaw Benchmark Shows Qwen3.5:27B Outperforms Other Local LLMs for Agent Tasks
A benchmark of 7 local LLMs on 22 real agent tasks using OpenClaw found qwen3.5:27b-q4_K_M scored 59.4%, while the runner-up qwen3.5:35b scored only 23.2%. Most models couldn't find basic tools like email functions.
Privacy-First MCP Server Directory Launches with Documented Data Handling Policies
A new directory at toolora.dev/mcp-hub lists MCP servers with documented data handling policies, including local vs hosted classification, what data each tool transmits, and whether accounts are required. The creator also provides a browser test method to verify privacy claims.